ARTIFICIAL INTELLIGENCE 41

Automated Space Planning

Charles M. Eastman
Institute of Physical Plmming, Carnegie-Mellon University,
Pittsburgh, Pa. 15213

Recommended by B. Meltzer

ABSTRACr
This paper both reviews current procedures and introduces new ones for the automated
generation of two-dimensional arrangements. General properties of the task and some
sufficiency conditions for dealing with it are identified. The treatment of these properties in
existing programs are r~'iewed. The task is also organized into its component decision rules.
One exemplifwation of these rules is described wWlichutilizes the sa/flciency conditions and is
implenwnted in the General Space Planner (GSP ~17rooram in operation at Curnegie-,~lellon
Unive.rsit.v [3], [4]. The performance of GSP in solvi.q9 a set of spatial arrangement tasks is
described and some future extensions outlined.
A secondary purpose of this paper is to more fully introduce this problem domain to the
artificial intelligence literature. Not only is it on inlerestino problem class now only the
province of humans, but it has wide application. Throughout the presentation, both the
commonalities and disparities of this task domain #~th other A! tasks are explicated.

1. Introduction
In this paper, i focus on computer programs allowing automatic resolution
of problems which humans solve using orthographic drawing. Typical
problems include the layout of floor plans, the arrangement of equipment
in rooms, site planning, and other forms of two-dimensional design ,asks.
In such problems, distance, adjacency, and other functions of arrangement
are a principal concern. To distinguish arrangement problems from otbers
in design, we call them space planning problems.
A variety of computer data structures have been developed which can
represent space and objects [5, 9, 17, 18], and operations have been imple-
mented for manipulating and testing the resulting arrangements. While it is
straightforward to use such systems in an interactive design mode, the
benefit and challenge of such programs comes from their potential for auto-
mated space planning, for automatically ~,ombining manipulation and test
operations in such a way that arrangements are quickly generated that meet
some predefined criterio,1. Several programs are currently running or are
Artificial Intelligence 4 (1973), 41--64
Copyright © 1973 by North-Holland Publishing Company
42 CHARLES M. EASTMAN

under development for automated space planning [4, 8. 10, 12, 17, 21]. Their
rationale is not to eliminate man from design problem solving, but rather to
eliminate from his responsibility those tasks which ate routine and uncreative.
It is anticipated that automated space planning will be a subsystem of
larger computer-aided design systems. It may be used:
!. For giving first approximations to complex and ill-defined arrangement
problems, facilitating quick involvement of decisionmakers so they may
iteratively elaborate the task objectives;
2. For generating the layoltts of mechanical rooms, service cores in office
buildings, mechanical and electrical equipment, and other arrangements
where functional requirements predominate;
3. As a replacement for prototype plans for hospitals and health care
spaces, military command posts, etc. In place of a set of designs incorporating
"good practice", a program will be provided capable of realizing a (possibly
unique) plan, satisfying all criteria, given any context.
The objective is to allow the machine to exhibit the intelligence of a low
level draftsman whose actions can be monitored and intervened by a human
designer. To date the complex structure of most space planning tasks has
allowed the development of algorithms which do only poorly when compared
with an experienced human draftsman.

2. Problem Formation
In previously published studies of how humans solve space planning problems
[6, !1], it has been recognized that designers do not treat th,eir problems
within the traditional framework of mathematical optimization. While on
the one hand their concerns are usually complexly related and involve
multiple objectives, the small amount of information available about a
problem does not permit meaningful development of linear or nonlinear
weighting schemes for combining all objectives.
In a general way, their treatment of problems is most e~sily understood
in t~rms of constraints. The total set of constraints defines a feasible solution
domain. The task is to find any solution within this domain. Optimization
of a sort does take place during iteration of the search sequence. When the
initial set of constraints results in a trivial search, either new constraints are
added or old constraints are replaced with new more restrictive ones. Con-
versely, when a current problem seems intractable, the constraints are often
relaxed. Optimization takes the form of ite:atively modifying the prcblem
definition until an appropriate balance between tractability and quality of
results is achieved. Problem solving effort thus plays an important role in
determining the final problem formulation.
Implicit recognition of the above procedure of working has led the designers
Artificial Intelligence 4 (1973), 41-64
AUTOMAI'FD SPACE PLANNING 43

of all existing automated space planning programs to rely on a similar

formulation. Some of the current programs also facilitate iterative definition
and resolution of a problem.
The development of the problem formulation requires some initial defini-
tions. A Space shall be a bounded or unbounded spatial domain in which the
spatial arrangement is to be developed. It usually consists of a room, bailding
shell or site or sheet of material. Design Units (DUs) are the physical
elements that are arranged on (or in) the Space. Each DU and Space occupy
a domain of fixed or variable shape and size. The spatial relations that are
to be satisfied by any arrangement are called S-Relations. They define the
constraints to be satisfied. The formulation is
GIVEN:
a a Space;
{bl, b 2 , . . . , bM} a set of DUs to arrange in that Space;
{ c , , c , . . . . ,cN} a set of S-Relations that must be satisfied (1)
for any acceptable arrangement;
{d,, d,,..., alp} a set of operators for manipulating the location of
DUs within the Space;
eo some initial design condition (which simply may be a),
FIND:
a set of transformations of the form
(e', bt-l, d `+~) ~ d +1, so that a state e' is generated,
such that
e~~ { c ~ , c2, • • . , c ~ } ,
where o represents a matching operation.
We call this formulation the Single Level Space Planning Task. Inspection
shows that eq. (I) depicts a heuristic search task that can be represented by
the unbiquitous OR-tree. Each node depicts a DU in a particular location.
Levels in the tree correspond to the sequence in which DUs are added or
changed; the first level corresponds to the first DU manipulated, etc. The
objective is to prove existence of a branch with particular properties. The
termination test is explicitly nrovided in the form of the properties c~, c,,...,c~,.
In the general case, any logical combination of DUs or S-Relations, e.g. the
conjunction of disjunctive sets, should be possible. In practice, most programs
are limited to conjunctive sets of both DUs and S-Relations.
An important characteristic of space planning problems is that the opera-
tors d~, d 2 , . . . , d, are not well-defined; many different operators may be
devised to generate locations of DUs. Most of the programs for automated
space planning have relied on a different set. Another characteristic is that
the initial node e° is arbitrary; it may be simply the Space a or the Space
with an arbitrary arrangement.
Artiflciol Intelligence 4 (1973), 41-64
44 C H A R L E S M. EASTMAN

T h e a b o ~ e f o r m u l a t i o n can depict m a n y typical d r a f t i n g problems. A n

e x a m p l e is s h o w n in Fig. 1. It d e s c r i b e s a typical m e c h a n i c a l r o o m l a y o u t
p r o b l e m . T h e D U s to be included in the Space a n d the set o f S - R e l a t i o n s to be
satisfied are s h o w n at the b o t t o m o f t h e figure.

• ' ,, '-- I I I .J..i

~r- ~ RETU.N IF.X~UST~,AN I r I
EXHAUST '%. AIR i I $U~5.. v

-
HANDLER~ MIXER
;
~ ICOOLER
nq,j,,,
HEATER I FAN

TYPICAL AIR~:ONDITIONING SYSTEM DESIGN

I- "~ ] 1
8. EXHAUST
FAN

t
I. 2{, ,
I
n
[ JL . . . . . ,
I I J
t_ -J $. COOLING 7. RETURN AIR
6. AIt MIXER UNIT
2 1t l " 0123 5
i i t t |

/ FRESH (ee¢
/ AIR
HANDLER
I
I I
I
I 3. ELBOW
I L
L__A 2. SUPPLY 4. HEATER
FAN

REQUIRED RELATIONS
i. DU3 adjacent tosid¢ I of DU2. 8. DU6 adjacent to side 2 of DUI.
2. Distance between Supply and pt. I. on 9. DU7 adjacent to side 2 of DU8.
DU2 < 10. 10. Distance between Air Return a n d
3. DU3 adjacent to side 2 of DU4. pt. ! on DUB < 4.
4. Dbtance between pt. 1 on DU4 and 1 I. Sightline to be free between door
pt. 2 on DU5 < 2. and pt. 3 on DU6.
5. Distance between pt. I on DU5 and 12. Distance between pt. 2 on DU4 and
pt. 2 on DU6 < I. any pt. can DU3 < 3.
6. Distance between Fresh Air .nlet and 13. Distance between pt. 2 on DU2 and
pt. 1 on DUI < 3. any pt. on DU3 < 2.
7. DU7 adjacent to side 4 of DU6.

FiG. I. An example of a space planning problem from the area of building engineering.
Artificial Intelligence 4 (1973), 41--64
AUTOMATED SPACE PLANNING 4~

3. The Representation of Objects and Relatiom

A variety of techniques have been developed for representing the two-
dimensional shapes of DUs and Spaces. A useful measure for evaluating
them is the resolution of shape allowed in relation to storage required [5, 18].
The constraint that points in space may be occupied by at ~nost one solid
has received special attention. Some representations implicitly exclude
arrangements not satisfying this condition. They do so by depicting empty
as well as filled space and only allowing new placements that fall entirely
within empty domains. Others treat overlaps as a type of S-Relation to be
eliminated by testing of an arrangement. Some allow a mixed logic that
includes en~pty, solid, and occasionally occupied DUs. Fig. 2 presents the

Solids Use Si~e Empty Space

Solids I N N Y
Use ,~l~ace [. . . . N Y Y
Empty Space ] Y r i r

FIo. 2. Overlap condition~ allowed for a three, valued space definition.

truth table for those combinations [3, 17]. In general, the points within a
shape are assumed to be homogeneous. It is desirable, though, tc reference
special points, e.g. a cable connection on a core memory box, and side of a
DU or Space, e.g. the front of a display unit.
Any particular problem is specified in terms of a Space and a set of DUs,
plus a set of S-Relations between them to be satisfied. All space planning
prograws treat S-Relations as boolean functions. An unresolved issue is the
definition of a general set of S-Relations with appropriate predicates capable
of depicting the conditions desired in a wide class of space planning problems.
Sets have been proposed capable of depicting a meaningful subset of problems
concerning room arrangement [9, 10], equipment arrangement within a
room [3, 17, 21] and building arrangements on a site [10]. As a specific
example, the S-Relations implemented in the author's GSP program are
presented in Fig. 3.
For ease of description, we say that an S-Relation which has as one of its
predicates a DU,,, "belongs" to DU,,,. In the following development, we
denote the predicates of an S-Relation with superscriW.s. That is, ~ depicts
DEw,, as belonging to S-Relation n. Most S-Relations implemented in
programs thus far belong to two DUs.
The reason space planning is inherently difficult is the large (potentially
infinite) number of location and orientation combinations that are available
for any single DU. The difficulty involved in evaluating I('cations for a single
Artificial intelh'gence 4 (1973), 41-64
4
46 CHARLES M. EASTMAN

DU is severely compounded by the interdependencies among DUs imposed

by their shapes as well as by the S-Relations. The exhaustive enumeration
of all combinations of locations for each DU is impractical for all but the
most trivial arrangement tasks.

S-RELATIONS:
(I) ADJACENT (A, B, SB)--This S-Relation specifies that element B must have all of its
side SB adjacent to element A. (A may aJso be a Space.)
(2) SIGHT (A, B, irrA, PTB)--This S-Relation specifies that Point PTA on clement A
be visible from Point PTB on element B; no solids may lie in the space directly between
them. Again, the points are optional and if not defined, all of the element facing the
other element must he visible;
(3) DISTANCE (A, B, F, PTA, FIB)--This S-Relation defines that Point PTA on element
A be less than F units away from Point F I B on element B. PTA and F I B are both
option~J parameters, if either or both are omitted, distance ~ measured from the closest
edge of:he element;
(4) ORIENT (A, B, SB)--This relation defines required orientations between elements
B and A. Side SB of clement B must face cletv.ent A.

LOCATION OPERATORS:
SCAN (A, B)--Sequentially generates the complete set of locations for element B, identified
by the projection of edges in the space A. (See. Fig. 6.) It sequentially considers place-
merit of the element in each; all four orientations of the element are considered.
SCAN is the most general location operator available in GSP;
PERIMETER (A, B, SB)---This operator sequentially generates the significant locations for
element B along the boundary of a specified and located element in the space, denoted
by A. The side SB of element B !s placed along the common border. One or more calls
of this operator can satisfy the ADJACENT S-Relation.
DISTLOC (A, B. F, PTA, PTB)--This operator .sequentially generates all locations satisfy-
ing the S-Relations DISTANCE (A, B, F, PTA0 FIB).

FK~. 3. The above S-Relations and Operators are those implemented in the current v e ~ o n
of GSP.

3.1. Loam Opmt m

In space planning, like all types of search, the types of S-Relations (ends)
that can he resolved and the characteristics of the operators available (means)
are intimately related. Let us consider for a moment the kinds of locations
that should be generated if different S-Relations are to be satisfied.
All S-Relations that have been incorporated into space planning programs
accept either a line of locations (generated along the perimeter of a DU, for
example) or a domain, possibly made up ef several disjoint subdomams,
in which any location L~equally acceptable. For example, the GSP sightline
relation between the point in Fig. 4 marked by an "x" and an unlocated DU
is depicted by the domain denoted by the gray cross-h,tching A. Any
location of a DU completely within this crosshatched area satisfies the sight-
line S-Relation. In the same way, the locations allowed by an adjacent
Artificial Intelligence 4 (1973), 41-64
AUTOMATED SPACE PLANNING 47

S-Relation with th~ rectangular DU at the top of the page are all within
the domain demarked by crosshatched area B. This same condition holds
for all S-Relations; they can be characterized by the domains they project.
{Some S-Relations also restrict orientation of a DU. Orientation is considered
in Section 4. I.)

FIo. 4. The areas projected by a sightline and an adjacent S-Relation.

Let us denote the domain in which locations are allowed by a S-Relation

n for a DU,, as s:'. That is
" {V ! e c, ,,
where L i.~ the set of all possible locations, s m is called the projection ot" c°'.
L for all but the ¢;udest metrics is very large. Consider now the domain
which is the intersection of a set ef projections, that is, ~ ~,s s~. This intersec-
tion can be found with one of the algorithms for finding the intersection of
a set ofdomains [!, 18]. The potential number of locations to be considered are
reduced to any within the intersection. The test of this .°-mailerarea is sufficient,
if a feasible location exists.
We call the above method for identifying constraint satisfying locations,
projective location generation. Projective location generation depicts the
relevant locations for a DU, but is limited by two conditions:
(i) It only allows consideration of S-Relations with already located DUs;
(2) the method assumes all currently placed DOs are located correctly.
The first condition requires that other" S-Relations may restrict locations to
a subdomain or the current intersection. The second condition requires
that automated :;pace planning involve iterative search procedures with
backtracking; no -equential set of subgoals based on DUs are likely to
exist which allow partitioning of the problem with complete independence.
No automated space planning system has been implemented since the
projective location generation technique was first introduced in [2]. They
all rely on a heuristic for generating trial locations. Below, the heuristics
Art~icial Imelligence4 (1973),41--64
48 CHARtFS M. EASTMAN

that have been used are interpreted and compared according to the insight
gained from projective location generation.
One approach used in existing automated space planning programs
responds to the interdependency of S-Relations and locations by binding
them together one-to-one, in a straightforward kind of means-ends analysis.
The IMAGE system, by Johnson and Weinzaplel at M.I.T. [10], starts with
an initial (arbitrary) arrangement and sequent/ally tests each DU for un-
satisfied S-Relations. Each failed S-Relation found evokes the operator
which finds a new location for a DU that satisfies the S-Relation. There
is one operator for each S-Relation. This approach simplifies the control
structure of a space planning system ~nd guarantees that each S-Relation
has at least one effective m e a n s for resronding to it. In IMAGE, multiple
unsatisfied S-Relations for a DU are treated by "averaging:" the locations
proposed by each S-Relation using a least-squares-means-fit procedure. The
new location is likely to be different from any of those that are generated
by one of the S-Relations and bears no correspondence with their inter-
section. This approach was adopted from Sutherland's SKETCHPAD
program [20], and used the same least-squares-means-fit to define the
shape of an element, when S-Relations were applied to the shape. But in
SKETCHPAD, S-Relations constrained the location of a single point,
e.g., defining the end of a line segment, rather than a location.
Grason has developed a program GRAMPA [8] which similarly binds
S-Relations to location operators. It incorporates operators which sequenti-
ally generate all locations satisfying any S-Relation. Iterated calls of an
operator alter the state de~ription so as to systematically search the projected
domain defined by an S-Relation. By relying on a unique dual-graph re-
presentation and a limited set of S-Relations, Grason was able to make each
S-Relation-operator pair contextually independent. That is, each operator
alters a disjunctive set of arrangement variables. By this partitioning of the
problem, Grason avoided the difficulties resulting from different operators
cycling on a set of alternative states. Convergence was assured. Similar
techniques have been developed independently by Krejcirik [12] and
Yessios [21].
Eastman [4] and Pfefferkorn [17], on the other hand, keep the binding
between S-Relations and location operators loose. In their programs, a
limited number of location operators are provided which serve to satisfy
a larger set of S-Relations. In Pfefferkorn's DIS program all the possible
locations for a DU are defined in one pass and stored on a list for evaluation.
A static evaluation is used for sequencing their further testing for overlap
conditions and S-Re'ations. A single location operator generates the list
and thus it is relied on for satisfying all S-Relations. The location set is
defined by the projected edges and perimeters of currently placed DUs
Artificial Intelligence 4 (1973), 41--64
AUTOMATED SPACE PLANNING 49

(see Fig. 6). Eastman uses the same basic location set, but his operators
sequentially produce them, one at a time, for evaluation. A set of operators
is provided, though, and some operators delimit the location set to those
which satisfy one of the S-Relations. Thus matching is possible between a set
of S-Relations belonging to a DU and the most appropriate location operator
for locating it. Eastman's and Grason's programs must try a fixed sequence
of locations before an acceptable one is found.
With these preliminaries, we may proceed with a more thorough discussion
of space planning algorithms.

4. Algorithms for Space Plaming

All automated space planning programs can be characterized by their
treatment of two component decision rules: (!) for selecting DU-operator
pairs to move from state to state; (2) for determining when each S-Relation
is applied to a state and the information gained as a result of that test.
Thus DU-operator pairs alter the current solution state; S-Relations generate
information about a state. Both rules can be elaborated. The first rule can
be considered as four subrules which distinguish the selection of Dus from the
selection of operators and by further distinguishing whether the just previous
tests passed or failed.
Several different starting conditions, e°, have been used or proposed and
each results in a different test interpretation. When the initial state is made
up of an arbitrary arrangement, only the failure of an S-Relation evokes
a tral~sition in the solution state. When all S-Relations pass, a solution is
indicated. The solution method resulting from this starting condition may
be called hill-climbing. Another starting condition is an empty space. This
starting state results irt a build-up solution strategy. DUs are added one at
a time, each DU placement being tested and possibly changed until the
S-Relations pass. Here, a passed set of S-Relations leads to the addition of
a new DU. A solution is indicated when all DUs are located. Other strategies
arc also possible, such as finding an arrangement in an unbounded Space
of two or more DUs related by S-Reiations, then locating these in the Space.
Here, S-Relations between certain pairs of DUs are resolved first and other
S-Relations resolved later. I call this a priority solution method. Mixes of
strategies are also possible. Under some conditions in all these methods,
a location operator may fail, indicating that the operator could not generate
an alternative location. This leads to the identification of another DU-
operator pair ~,=d in buile-up approaches the DU for which no acceptable
location is reraoved from the arrangement.
Most of the programs to date sequentially add DUs to an arrangement.
(Both build-up and priority strategies sequentially add DUs.) The principal
Arti/icial in,'elligence 4 (1973), 41-64
50 CHARLES M. EASTMAN

advantage of sequentially adding DUs derives from the condition that only
corr~tly placed DUs contribute to the search of locations for other DUs.
Incorrectly placed DUs consume space and restrict placements in arbitrary
ways. The only potential benefit from hillclimbing would derive from the
capability of abstracting useful information from incorrectly placed DUs.
No proposals on how to do this have been put forward.
With these points in mind, I now turn to the presentation of decision rules
based on projective location generation for all the cases outlined above.
In most cases, the rules presented have been implemented in the GSP program
now in operation at Carnegie-Mellon University and elsewhere. Also included
in this section are the decision rules utilized by the other automated space
planning programs now in operation. / The specific implementation of the
decision rules in GSP is presented in Section 5 and their performance outlined
in 5. I. Some extensions are suggested in Section 5.2,

4.1. Toting of States with S-Rdations

A review of the S-Relations used in all space planning programs shows that
there are two types. The first type c,f S-Relation is only affected by the
placement of the two DUs it belongs to. An adjacency S-Relation between
two DUs for instance, will not t e affected by the placement of other DUs.
We call this type of S-Relatior: local. Alternatively, a sightline relation
between two DUs may fail because of the placement of the two DUs it
belongs to and also from an interfering location of any other DU. We call
these S-Relations global.
Consider the following rule. "All S-Relations belonging to a DU must be
applied when the location of that DU is being evaluated." This rule becomes
valid only if we consider global S-Relations as belonging to all DUs. Johnson's
IMAGE program [10] explicitly validates this rule by giving to all global
S-Relations a predicate identifying the global set of DUs. Eastman's GSP
implicitly incorporates the rule by testing the location of any DU with those
local S-Relations belonging to the DU and also all global S-Relations.
Pfefferkorn's DPS tests sightline S-Relations after the placement of every
DU and other S-Relations ifthey are predicates to the object [17]. Yessios and
Grason use only local S-Relations.
Local S-Relations require less testing than global ones. Another way of
reducing S-Relations is by logical elimination. Those S-Relations belonging
to a DU whos,-, projected domains subsume the domain projected by another
S-Relation are "slack" and have no effect in determining the final location,
~The breakdown of the remaining programs are as follows: Grason's program [8]
relies on an exhaustive lexicographic search. Se:~ral others, such as Maver's [13], N e ~ o -
ponte's (15], and Newman's [16] operate in an interactive mode only, without solution
algorithms.
Artificial Intelligence 4 (1973), 41-64
AUTOMI rED SPACE PLANNING 51

If they may be identified, they can be eliminated. For example, with the
appropriate definition of adjacent and distance S-Relations, an adjacent
S-Relation between two DUs subsumes all distance S-Relations between
them. Other S-Relations partiaily define a location in such a way that
inconsistent S-Relations may be identified. For example, orientation is not
considered in the projective location generation technique. Testing for
inconsistency of orientation restrictions among a set of S-Relations is easily
implememed and applicable to sightline, adjacency, and orientation and other
S-Relations. Montonari has initiated other work in constraint elimination
which eventually may be applicable to space planning [14].
In review, after logical elimination oi all possible S-Relations, those
S-Relations relevant to any DU placement arc those which we have defined
as belonging to that DU.

4.2. The sequencing of S-Rel=tions

In most automated space planning programs, tests are sequentially ordered
according to the DU they belong to. That is, all S-Relations belonging to
one DU are tested, then those ofanother DU, etc. We call this testin9 by DU.
One reason for testing by DU is the above rule regarding local and global
S-Relations. Elimination of some testing is allow~ when S-Relations are
ordered by DU. It is net necessary to order S-Relations in this way; the
orientation of all DUs could first be tested, followed by testing of their
placement in the x dimension, etc. This alternative mode of sequential
testing by property is used by Grason [8], Krejcirik [12], and Yessios [21].
The rationale for either sequence is to group tests in such a way that ~he
maximum information about a variable of the alternative s~ate may be derived
to determine if and what kind of action is required.
Whether S-Relatior.s are ordered by DU or by Property, further sequencing
is involved in applying multiple tests t c a single DU or property. Some
sequences are more effcient than others. Sequences of S-Relations are also
required to derive the locatio~l of a DU using projective location generation.
That is, all algorithms for finding the intersection of domains [7], [18] find
the intersection of only two at a time, e.g., (. • • ((st c~ s,) c~ s 3 ) " • ~ st).
The intersection operation, like all S-Relations, can be considered a con-
junctive boolean test; the halting condition is complete passage or any
failure. =
The most efficient sequence for testing conjunctive boolean tests [19] is in
ascending order of
cost of executing the test
probability of failure (2)
= T h e halting conditions are {V c,~l cn = T R U E } v I] c,,I c~ = FALSE}.
Artificial lmeiligenc¢ 4 (I 973), 41-64
52 CHARLES M. EASTMAN

The probability of failure for the intersection operation on s," or the S-

Relation ~ may be assumed to be inversely proportional to the area of s,',
denoted SA, divided by the area of DU,, denoted A ' . This important
assumption proposes that for the intersection operation, the likelihood
of a null intersection is inversely proportional to s~ (divided by a constant).
For the S-Relations, the ass~,mption is that the likelihood of failure is inve~zely
proportional to the number of locations allowed by the S-Relation and that
the number of different locations allowed by an S.Relation is proportional
to the area covered by its projection divided by the area of the DU. Thus,
if we denote the cost of executing c,, as R., then eq. (2) is applicable to space
planning in the form:
R.SA'~ (3)
/l~'= Am "
That is, S-Relations with smaller projected areas which can be quickly
executed should be considered first. Correspondingly, if R,, is the cost of
executing the intersection operation on SA,, then eq. (3) characterizes the
sequence in which operations are to be implemented. It is straight-
forward to estimate by a simple function the projected area SA," generated
by c~',; see Fig. 5. These functions can be used with eq. (3) to order S-Relations
such that their order is 17,,j ~</1=2 ~ • • • 13~, k ¢m. The intersection operation
is executed in the order (. • ((s, n s2) c~ s3)- • • c~ s~). The same conditions
hold for testing an existing arrangement, as in hill-climbing. The S-Relations
should be considered in increasing order of their l~-values.
Even without location operators utilizing projective location generation,
the above rules are applicable, if more than one location operator is available,
select that one that responds to the most restrictive S-Relation, that is,
dp - {cvl/1,1, where n = p}. in this case, though, the cost R is the cost of
locating DU,,. The locations generated are tested with the other S-Relations,
ordered in the ascending sequence generated by eq. (3).
From the insights available from projective location generation, we have
seen that readily available information ,'egarding S-Relations can be used to
sequence location operations and testing of a DU in an efficient order.
Only Sutherland has considered the test ordering problem for automated
space planning [20]. The methods used in SKETCHPAD approximate
those implemented in GSP, as described later. The function SKETCHPAD
uses o order tests has not been published. Other programs have not considered
~he test ordering issue and have taken them in arbitrary order.

4.3. State Transformations

We have already considered how a sequence of S-Relations should be ordered
for maximum efficiency. A similar rule can be developed for sequentially
Artificial lntelilge~ce 4 (I o73), 41-64
AUTOMATED SPACE PLANNING 53

adding D U s to an arrangement. Recognizing t.at each D U has associated

with it a set of S-Relations, we may order the D U s by ordering a functional
characterization of their sets.

in general:
A m --: area o f D U m
As = area o f the space
A*i = area available for locating DUt
= A . ' - ~ ,4,- ~ : . A s'
x m a n d .ym -= lenglh and width o f DUm, respectively
Wam __ width o f side a o f DUS"

F o r A D J A C E N T (A, B, SB):
sa~ej: = ( ~ + xb - , ~ ) " (x- + >..)" J.5 • a.__b
As '
A*a
S J:dj: + - ( 2 * A ' - - (.P~*Z))4" A---;'

where z: -~ { ; : = ~_, 2w~llw~'

l > ~}.
tfm

For SIGHT (A. B, PTA, PTB):

S,~.~,: = ( a *° - ~" (x+ + .v+)) * ~ * A*o
• tim As "

A*b
magi,: = (A'b -- 7 (xl + yl)) * ~ * A, "
~m
where i = V~, i :~ a, i :~ b,
a = 1.5 i f P T A ~ 0 A P T B ~ 0 ,
ct --- 0.8 if PTA ~- 0 V P T B ~ O,
0t = 0.3 i f P T A = 0 A i r r B = 0.

For DISTANCE (A, B, D, PTA, PTB):

Aea f 0 if PTB = 0,
S,4~ist: = ( d + ~ -t- ~ . ~ ' , -~;-, whcgl¢
0t = [~(xb + .+4,) otherwise

=
{i(xa + ~<) : PTA #- 0,
t(xa + ya) otherwise.
{4 i, ~ #= space, a:~spaoe,
Y= otherwise.
Aoh
S a l t , , : =. ( d + o + ~).o* As "
(with aq a and h scripts reversed).

For ORIENT (A, B, SB):

A*b 4- A*b
SA~ort: ----
As

FIG. 5. Functions for estimating the area allowed by an S-Relation. In all cases A'liAr is
the t~roportion o f any projected area likely to be filled with D U s other than i.
Artificial lntelliRence 4 (1973), 41--64
54 CHARLES M. EASTMAN

In lieu of actually projecting the domains, defining their domain of inter-

section and its area, the area can be estimated directly from the areas of the
single projected domains, given in Fig. 5. It is necessary to assume independ-
ence among the projected areas, that is, that the likelihood of any point in
the Space falling in two or more domains is the product of the likelihoods
that it will fall in each one of the domains. This assumption can be made
only after the logical elimination of S-Relations has been attempted. The
likelihood that any x - y point will lie in the projected domain of one S-
Relation n is SA~/TA ~ where TA m is the empty space available for DU,,.
The area of the intersection of the set of projected domains for DUm is
SA," (4)
TAM I'I TA m.
hen
Combining this with eq. (3) gives
R,TA m SA~
.4- 11 TA''
R,N
(5)
where R, is the cost of locating DUm. DUs added to the arrangement in
ascending order ofeq. (5) will result in an efficient sequence. But it is not likely
to be optimally ~fficient, for the following reason. As DUs are add~l, the
3-Relations which belong to located DUs change. Thus adding a DU to the
arrangement adds S-Relations to the sets of other DUs, possibly changing
the order. No sequence of DUs may actually result in an ascending order.
Equation (5) derives a single value that is a function of both a DU and the
operator which locates it. Both are given consideration in determining the
next operation. Rules for these two aspects of the decision have also been
derived independently; see [4]. Pfefferkorn's DPS approximates the above
sequence by associat,ng with each S-Relation type a severity measure. DPS
sums the severity measures for each DU and adds to the arrangement next
that DU which has the largest value, e.g., most severely constrained. Ties
are broken by taking the largest DU first [17]. Other space planners locate
DUs in the order given by the user.
A requisite of any location operator is that it determine if an intersection
exists of the projections of the relevant S-Relations and if the intersection
includes an empty space large enough to hold the DU. if either of these
,:onditions does not hold, the operator will fail. Failure is caused by ~ny of
three different conditions:
(!) A set of S-Relations that, when their domains are projected from the
current arrangement of DUs, result in a null or improperly formed inter-
section.
(2) An intersection exists, but it is blocked by a D U being improperly
placed, that is,
(a) some DU has been placed in the projected intersection originally
Artificial Intelligence 4 (i 973), 4 !-64
AUTOMATED SPACE PLAN~NING 55

defined for it and also within the intersection defined for the current
DU, or
(b) the projection of a global S-Relation is being altered by the location of
some already located DU.
Alternatively, it may be that
(3) No arrangement exists that satisfies all S-Relations.
The first two failurc conditions can be remedied with appropriate responses.
To date, no test of necessary conditions have been devised that call distinguish
these first two from the last.
Insight into the above failure situations may be gained by imagining
implementation of projective location generation on a graphic terminal.
The current arrangement is shown on the screen, the DU-operator pair is
selected for addition and the proj~ted domains for the S-Relations belonging
to the DU and to located DUs are projected simultaneously on the screen.
Each domain is crosshatched, if only one domain does not intersect with the
others, the DU to relocate would be easily identified; it would be that one
which belonged to the S-Relation whose projection did not overlap the others.
Similarly, if an intersection did exist but was already occupied by an already
located DU, or if a projection was blocked by a DU, it could be seen. The
area coming closest to being an intersection would be brighter due to multiple
crosshatches. The DU causing non-overlapping should be moved so that
its projected area overlapped the bright area and so that the DU's domain
did not interfere with the intersection. If multiple areas did not overlap,
the combinatorial search becomes larger, but the insight provided is still
,.'onsiderable.
Recall that intersections are derived from an operator which takes two
predicates at a time. A failure may occur between the current intersection
and any new domain. Thus, if a null intersection occurs, no straightfor¥~ard
means exists to derive the partial intersections, as was done visually by the
human eye on the graphic terminal. No efficient procedure for summing
the number of proje~ions which cover each x - y point has yet been proposed.
Given these difficulties, a man-machine system relying on man's pattern
recognition capability is one resolution. For completely automated space
planning, several backup nlles for re-locating DUs have been utilized whic'.,
respond to the conditions just depicted. The first relies on the probability
of failure of any S-Relation. Given a set of S-Relations "linking" one DU
to a set of DUs, the S-Relation most likely to not cover any particular set
of points is that one covel'ing the smallest area; thus if m* denotes the s~V-
script o f a j u s t failed DU and m' deaotes an already located DO, then relocatt;
D U , ' such that
rain SA ''*m'.
Another response, is to relocate any DU which is currently located in an
Artificial Intelligence 4 (1973), 41-.64
56 CHARLES M. EASTMAN

area which is in the intersection of the projected areas for the failed DU.
These two rules approximate the graphic resolution proposed above.
Pfefferkorn uses two other techniques, if only one S-Relation has continually
failed (e.g., is the only one not in the intersection) he removes the current DU
and the located one which belongs to '.he failed S-Relation. An arrangement
of these two DUs is found in an unbounded space and that arrangement is
treated as a macro-DU and located, if possible, in the arrangement. That is,
he switches to a priority solution method (see Section 4). if this response.
is not applicable or does not resolve the failed condition, the problem is
started again but with the failed DU placed ":t the arrangement earlier [17,
p. 120]. Johnson and Weinzaptel's program [1 J] does not search the possible
locations for a DU and thus cannot determine if a DU being considered has
a location satisfying all S-Relations or not (see Section 3). When more than
one DU has S-Relations associated with it which fail, their program rel,- ;ates
that DU which has the largest number of failed S-Relations associated with it.
(They also weight S-Relations by importance. In this case they sum the
weights of the failed S-Relations).

5. The GSP Implementation

GSP utilizes a build-up strategy and incorporates many of the rules previously
described. In general, its search behavior can be described according to the
three decision rules inherent to the build-up approach. They are:
1. the sequence in which DUs are added to the arrangement and the
operator used for the addition;
2. the sequencing of S-Relations used to evaluate a location;
3. when no location can be found for the current DU, the back-up rule
used to select relocations for already located DUs.
Like other automated space planners, GSP does not incorporate projective
location generation. It relies on a heuristic set of possible locations, with
testing to eliminate all but the acceptable ones. Specifically, the basic set
of locations considered by the system are the four 90 ° orientations of a DU
in each corner defined by lines projected by the edges of the Space and all
located DUs; see Fig. 6. Each of the twenty-four locations shown would be
considered four times, once for each orientation of the DU. The most general
location operator generates this set, one location at a time, in the sequence
sho~rn. Two other location operators generate subsets of these location'.:,
in different sequences. The first generates the subset that is within the area
projected by the GSP adjacency S-Relation between the to-he located DU
and any already located DU or the Space (refer to Fig. 3 for the description
of the adjacency S-Relation). The last generates those locations within any
area projected by a specific distance S-Relation. GSP must select the appro-
Artificial Intelligence 4 O973), 41--64
AUTOMATED SPACE PLANNING ~7

priate operator for generating the initial ~ t of locations, then sequence the
relevant set of S-Relation *.o each trial location. Currently, GSP is limited to
arranging DUs of fixed size and shape.
For each S-Relation, GSP uses the functions in Fig. 5 to estimate its
projected area. These estimate~ make assumptions about shapes (that all are
rectangular) and the distribution of elements. They are static, in that they
rely on the particular problem definition and not on any particular arrange-
ment. (We expect to improve these methods for estimating areas as our
experience with GSP grows.)

F
a13"

9
~T-- -~
¢ls

I0
j
12
7 II I2i"
t - - - G ~0

FIG. 6. in the twenty-four locations shown, each elcmcr,t would be considered in its four
orientations. PERIMETER (in relation to the space) would generate 1-2-3-4-$-6-12-fi-
16-20-22-24-23-17-13-7 in that order.

We have found that incorporation of the previously described decision

rules is facilitated if the information they utilize is represented in the form of
a directed graph. Each DU and the Space is represented as a node in the
graph. Each S-Relation is represented as one or two directed edges. If an
S-Relation is unary or is between an element and the Space, then only one
edge is defined. Each edge has a value corresponding to the area allowed for
its successor element when its predecessor element is already located. Each
node then takes one of three values denoting whether it is unlocated, located,
or immovable. For example, the problem shown in Fig. 1 can be represented
by tire graph shown in Fig. 7. We call such a graph a Constraint Graph.

FK~. 7. The Conctraint Graph for the problem shown in Fig. I.

Artificial latelliegence4 (1973), 41--64
58 CHARLES M. EASTMAN

The Constraint Graph is stored in GSP as a list of directed edges in array

T., where Tni is the predecessor node, 7",, is the successor node, T,3 is the
value of the edge from Trot to T,z, and T o is the cost cf locating a DU using
the most efficient operator associated with the S-Relation which T,, depicts.
Also the Constraint Graph is ordered in ~zscending order according to the
value of T,3. That is T i j <~ T,2.3 ~< " ' " TN3. For each of the M DUs we
store a value a,, m - 1, 2, . . . . M. 0t gives the status of each DU, where
=,, = 0 indicates that DU= is unlocated, 0~,.- I indicates that DU,. is
located but removable, and 0t~, = 2 indicates that DU,, is located in a fixed
location, e.g. a door in a room that cannot be relocated.
Prior to each new DU being added to the arrangement, the value of T,, is
determined for each unlocated DU. 7",4 is assigned a cost value equal to the
smallest R , associated with the S-Relations which currently be!ong to the
DU. When a DU is added to the arrangement, that DU is chosen for which
the following function is minimized:

T,.,
rain T,~TA'" IF[ TA~,, where {i ¢ Ni %-, = (l v 2) ^ ( ~ , , - 0)}. (6)
Am i, M
A" is the area of DU,, and T 4 " is the total space available for locating DU,,.
In general,
M
TA m = A° + A" - Y" Ak where A° is the area of the Space a.
It,,I
Eq. (6) corresponds to eq. (5). The operator whose cost i~ d :noted by 7~4 is
used to initially locate DU,. We den.:te the index of the DU just located
by an asterisk, that is m*, and previously located DUs by a prime, e.g. m'.
The sequence of testing is simply in order of T. That is, test S-Relation n if
{n E NI (T,I = m') ^ ((T,2 = m*) v (T,, = global))} (7)
in the order defined by n. Eq. (9) corresponds to eq. (3), where all R, are
assumed equal.
The back-up conditions are called into play whenever a location operator
fai!s to generate an acceptable location for a DU. Let us denote *lie subscript
of the failed DU as m•. The DU to relocate is m', where
m" = {T,,I min T,3 ^ ~r,= = 1 A T,z = m~}. (8)
if no DUs satisfy this condition, a more complicated backup procedure is
invoked. Each of the locations of the locatcd DUs is considered as a location
for m~. Specifically, the set of S-Relations belonging to nP is tested by
replacing m* with the set of DUs satisfying the condition {m[ 0~, = !}. If
c, = 0, !, then
m" = {m"J (max Zj ~'l y~c~" >1 J)} (9)
where { j ¢ NI (Tj2 = nO) ^ (~rj, = ])}. Ifthe no DU m,~ts these conditions,
the space planner fails.
Artificial Intelligence 4 (1973), 41--64
AUTOMATED SPACE PLANNING ~9

The Conztraint Graph depicting any particular problem is not used as an

ahernative representation in which to search for a solution, but an informa-
tion structuring ~presentation that allows insights ~o be derived that will
facilitate search. The search of arrangements uses a variable domain array
spatial representation [6]. The Constraint Graph depicts in an explicit way
the relations existing in a space planning task. It is used in a manne~ similar
to Gelerntcr's planning model for geometry theorem proving. Johnson and
Weinzapfel [10] first structured S-Relations in a graph form, though they
did not utilize it to se'.luence all decisions. We ~all space planning w~lich
structures decisionmaking according to a Constrain t. Graph Constraint
Structured Planning.

5.1. Performance
GSP is a program of about 5500 card images written in FORTRAN IV.
It requires 35 K words of core memory and runs on Carnc£ie-Mellon
University's IBM 360, Model 67 under the TSS timesharing monitor. It has
been given a variety of problems, using both Constraint Structured Planning
and less considered decision rules (for an alternative set, see [3]). Below, its
performance in solving four small scale problems is presented.
Three different measures of performance are provid~,d. The first measure
is the number of state transitions required to transl~,te an initial problem
condition into an arrangement satisfying the specified S-Relations. A state

o 12. 4:]

, !
|._-.

ADJACENT(I I, I, 4)
ADJACENT(! I, 2, I)
S[GHT(I !. 2, I, 3)
--SIGHTO I, 4, I,o)
DIST(2. !. 12. I. 1)
DIST(2, :3, 2, 2, I)

]" I ul

FK;. 8. Small mechanical room.

Arti.~ciallnlelBtlence 4 0973), 41--64
 60 CHARLES M.

transition is one placement of a DU. The number of state transitions to solve

a problem is a measure of the search efficiency of the decision rules used.
A lower bound for state transitions is the number o f DUs. The second
measure is the C P U time rf;quired for specifying the task to GSP and for it
to solve the problem on the above machine. This measure depicts the efficiency
of tbe system from an economics and practicality standpoint. The last
measure is CPU time per state transition. This measure characterizes the
efficiency of implementation of the mix of representation, S-Relations and
operators.

r w
I-,d

DISTO, 5, 40, I, !) • ADJACENT(II,6, I)

DIST(5, 6, 20, I, I) DIS'TO, 4, 45, 1, I)
ADJACENT(I I, 3, I) ADJACENT(I 1, 1, I)
ADJACENT(I I, 5, 1) SIGHTO, 2, I. 1)

F' "1
Fxo. 9. Small oflk¢.

The first three problems and the solutions found are shown in Figs. 8, 9, 1O.
The last problem is the original one specified in Fig. 1. Tl:.e sequence of states
generated in finding a solution to this last problem is shown in Fig. I I. The
measures of p e r f o r m a n ~ for each of these problems is shown in Fig. 12.

SIT(I 1, 1, 1, 2)
SIT(I 1, 2, 1, 2)
s ADJ(II, I, 1) --SIT(I I, 3, I, Uj
ADJ0 I, 2, l) DIST(2. 3, 6, 1, I)
~ ADJ(II, 3, 1) DIST(I, 6, 6, 1, 1)
I I ADJ(I I, 4~ 1)
ADJ(i 1, 5, 1)
ADJ01.6. 1)

Fro. 10. Medium machine room.

Aru'flcial Intelligence 4 (1973), 41-64
AUTOMATED SPACE PLANNING 61

!r

Fro. 1I, State transitions for the problem in Fig. 1.

Its general performance indicates that Constraint Structured Planning

works most efficiently when the task is highly structured. The search space
is made significantly smaller if many adjacency or distance S-Relations are
involved. If no S-Relations project areas for unlocated elements, they are
chosen according to their approximate size, biggest first. GSP is knov,r, to
fail or o',tke extremely poor progress on some problems, due to poor handling
of global S-Relations. More efficient coding could significantly lower current
running times.

No. of CPU time Average

Problem No. of DUs S-Relations States (seconds) seconds/state

Problem Eight 6 4 IO.S5 2.6

Problem Nine 8 55 144.40 2.6
Problem Ten 13 8 l 1.10 1.4
Prob'~m in Fig. l 13 86 101.47 1.2

Fro. 12.
Arti/icial lntelltge~e 4 (1973), 41-64
5
 62 CR~L~S M. ~ s r ~ N

5.2. Exteuslom
While Constraint Structured Planning, as implemented in GSP, begins to
effectively dez.! with several of the issues inherent in space p!ahning, many
remain for more detailed exploration. Current efforts at expanding GSP
include the imroduction of variables shape DUs and the appropriate opera-
tors, S-Relations and control strategies to deal with them. GSP is also being
extended so as to handle hierarchical problem domains, e.g., "arra:tge the
equipment in the rooms, which are arranged and shaped within a building,
which is arranged and shaped on a site." Such formulation~ allow precise
study of problem decomposition in an interesting comcxt. And of course,
we have yet to take full advantage of projective location generation.

6. CondusioR
Automated space planning is an excellent example of a search task; the
interrelationship~ of variables resulting from DU shapes and S-Relations
lead to problems intractable using standard mathematical approach,-s. No
representation nor partitioning of variables has yet allowed effective circum-
vention of a process of heuristic search.
I have attempted to indicate the task structure and some efficient decision
strategies utilizing that structure. Projective location generation greatly
delimits the locations to be considered for a DU. It provides an operator which
is sufficient for finding a location, if the conditions of the operation are met.
Constraint Structured Planning utilizes the structure provided by S-Relations
to efficienOy sequence and select state transformations. It is a unique way
to relate the set of goals to be achieved (S-Relations) with variables in ".he
problem space (DUs) and to use the relations to sequence search operations.
The currently used analysis of the Constraint Graph is incomplete, though.
Only S.Relations which have their predicate elements located can currently
be brought into the analysis, yet the other S-Relations act to delimit allowed
areas also. No means has yet allowed us to estimate the restrictions imposed
by all S-Relations. Thus the most efficient sequence of DUs to add to an
arrangement cannot be dehned.
Certain underlying conditions in the problem domain warrant further study.
Within the problem domain considered here, is.;aes still remain in each of the
three decision processes. In the element selection process, backup conditions
are those where GSP is most likely to encounter difficulties. A better under-
standing of failure situations is still required. Necessary conditions to dis-
tinguish local t~tilures from infeasible solutions would eliminate much fruitless
search. Also the treatment of local S-Relations is much more efficient than
for global ones. There is some indication that it is possible to translate all
global S-Relations into one or a set of local S-Relations. In GSP, estimations
Artificial lnt elllL,ence 4 (1973), 41-.64
AUTOMATED SPACE PLANNING 63

of the areas projected by ~ are used to determine the likelihood that S-

Relation n will fail. This is in comparison to actually projecting the area.
The estimating procedure becomes less reliable when the area of the inter-
section of a set of projected domains is derived. Comparison of Constraint
Structured Planning using estimated veto.us actual area of projected domains
is required to determine whether the assumptions of independence (Section
4.2) are reliable.
As greater capabilities are redeveloped for processing spatial arrangement
tasks, we may expect better ~.nalyzed, better resolved, and possibly even
more beautiful physical environments to result.

ACKNOWLEDGMENTS
A preliminary version of this paper was presented at the Second International Conference
on Artificial Intelligence, London, 1971. This work is supported by the National Science
Foundation, grant no. G J31188.

REFERENCES
I. Eastman, C. M. and Yessios, C. An efficient algorithm for finding the intersection,
union arm differences of two domains. Inst. of Phys. Planning, Res. Rep. No. 31,
Carnegie-Mellon Univ., August 1972.
2. Eastman, C. M. Heuristic algorithms for automated space planning. Proc. Second
Intern. Con~ on Artificial Intelligence, British Computer Society, London, 1971.
3. Eastman, C. M. Prelhninary report on a system for general space planning. Comm.
A C M 15 (2) (February 1972), 76-87.
4. Eastman, C. M. GSP: A system for computer assisted space planning. Proc. Design
Automation Workshop, June 28-30, 1971, Atlantic City, N.J.
5. Eastman, C. M. Representations for space planning. Comm. ACM 13 (4) (April 1970).
6. Eastman, C. M. Cognitive processes and ill-defined problems: case study from design.
Proc. Joint Intern. Conf. on Artificial Intelligence, D. Walker and L. Norton (eds.),
Washington, D.C., May 7-9, 1969.
7. Gelernter, H. Realization of a 8eometry-',hcoretr~ proving machine. Computers and
Thought, E. Feigenbaum and J. Feldman (eds.), McGraw-Hill, New York, 1963,
pp. 134-152.
8. Grason, John. An approach to computerized space planning usinff graph theory.
Proc. Design Autonmlion Workshop, June 28-30, 1971, Atlantic City, N.J.
9. Grason, John. A dual linear graph rep;escntation for space filling location problems
of the floorplan lype. Emergittfi Meti.ods of Environmental Design and Planning,
G. Moore (ed.), M.I.T. Press, Cambridge, Mass., 1970.
I0. Johnson, T., Weinzapfel, G. et ul. IMAGE: An interactive graphics.based computer
system for multi-constrained spatial synthesis. Rept. Dept. of Archi:ectute, M.I.T.,
Cambridge, Mass., September 1970.
11. Krauss, R. and Myer, J, et al. Design: A case history. Emerging Methods of Environ-
mental Design and Planning, G. Moore, (ed.), M.I.T. Press, Cambridge, Mass., 1970,
12. Krejcirik, M. Computer-aided plant layout, Computer Aided Desagn 2 (I), 1969.
13. Mayer, T. W. P A C E I : Computer aided building appraisal, Architects'J., July 28, 1971.
14. Montanari, V. Networks and constraints: Fundamental properties and applications
to picture processing, Dept. of Computer Science Working Paper, Carnegie-Mellon
Univ., January 1971.
Artificial Intelligence 4 (1973), 41-.64
gQ
64 ~ M . F.ASrMAN

15. Negroponte, N. and Grossier, L. URBAN 5: A machine that discusses urban design,
Emergino Methods of Envirommntal Dis!on and Piannlno, G. Moore (ed.), M.I.T. Press,
Cambridge, Mass., 1970.
16. Newnum, William M. An experimental program for architectural design, Computer J.
9 (June 1966), 21-26.
17. Pfefferkorn, Charles. Computer design of equipment layouts using the design problem
solver, Ph.D. Dissertation, Dept. of Computer Science, Carnegie-Mellon Univ.,
May 1971, unpublished.
18. Phaltz, John L. and Rosenfeld, A. Computer representation of planar regions by their
skel2tons, Comm. ACM 10 (2) (February :967), ! 19-123.
19. Slagie, James. An efl~'knt algorithm for finding certain minimum cost procedures for
making binary relations, J. ACM (1964), 253-264.
20. Sutherland, I. G. Sketchpad: A man-machine graphical communication system,
AFIPS Con~. Proc. 1963, Spartan Books, Washinston, D.C., 1963.
21. Yessios, C. I. FOSPLAN: A formal space planning language, EDRA FOUR: Proc.
Fourth Environm. Design Res. Assoc. Co~., W. Mitchell (ed.), UCLA, Los Angeles,
Calif., 1972.

Received 30 August 1972

Artificial Imelligence 4 (I 973), 41-64